Feature Selection Based on Layer-Wise Relevance Propagation for EEG-based MI classification

Brain-computer interface (BCI) enables communi-cations between humans and their surroundings via external devices. Electroencephalogram (EEG) is one of the common measurements of brain activities in BCI due to its non-invasive and relatively inexpensive characteristics. There have been a number of studies on the EEG-based motor imagery (MI) classification and its practical applications, including healthcare and rehabilitative technologies. Recent studies have applied deep learning techniques for MI classification, especially utilizing con-volutional neural networks because of their ability to generalize feature representations. However, it is very crucial to select only useful features in MI classification networks, because EEG having a low signal-to-noise value may include unnecessary features that interrupt the model decision. In this study, we perform feature selection in the extracted EEG features based on the layer-wise relevance propagation (LRP) method for MI classification. Our proposed framework is evaluated on the BCI Competition IV-2a dataset in the subject-dependent scenario, and our experimental results show that LRP-based feature selection improves the MI classification performance.